Reasoning block doesn't appear

Short answer

Not all models produce reasoning. Only reasoning-tuned models (DeepSeek R1, Qwen QwQ, Qwen3-thinking) emit <think> blocks. Regular models just answer directly - no block, and that’s normal.

Detailed explanation

Expected behavior: the “Thinking… (Ns)” card only appears if the model emits <think>...</think> in its output. This is a property of the model itself, not a setting in Abliteration.

Models that produce reasoning blocks:

• DeepSeek R1 (all sizes and distills)
• Qwen QwQ 32B
• Qwen3-thinking variants
• Some experimental fine-tunes

Models that do NOT produce reasoning blocks:

• Everything else - Llama 3.x, Qwen 2.5 (non-thinking), Mistral, Gemma, Phi, etc.

If you’re using a non-reasoning model and expecting a reasoning card, that’s a mismatch - pick a reasoning model instead.

If the model is a reasoning model but no card appears:

Rare but possible causes:

1. Model output doesn’t contain <think> tags. Some reasoning-model fine-tunes strip these before outputting. Check the raw response for <think>...</think> - if it’s missing, the model’s fine-tune removed it.

Fix: use an official version of the model instead of a random fine-tune.

2. Chat template misses the reasoning marker. If the model is expecting a specific system prompt or trigger to activate reasoning mode, and it’s not present, it may just answer directly without thinking.

Fix: check the model’s Hugging Face page for how it activates reasoning - usually just standard chat, but some require an explicit “Please think step by step” system prompt.

3. Model is confused about its role. Some reasoning models are sensitive to conversation history. If earlier turns pushed it out of reasoning mode, later turns may not think.

Fix: try a fresh chat with the same model.

Why it matters

The reasoning card is a visual affordance for a specific model behavior. Understanding that it depends on the model (not the app) saves debugging time.

In Abliteration specifically

The reasoning card is triggered when <think>...</think> appears in output. Detection is automatic - no toggle to turn on reasoning rendering. If tags aren’t in output, no card.

Reasoning is persisted to the database - if you saw the card during generation but not after chat reload, that’s a bug (please report). Normally reasoning survives reload just like regular messages.

Related

• Reasoning models - how reasoning rendering works
• Reasoning persistence - reasoning across chat reloads